[Hyacinth-of-Security (lxcxjxhx)]

What

Three text-mode open() calls in shipped code rely on the platform default encoding. On Windows that is the ANSI code page (GBK on a zh-CN machine, cp1252 on en-US), so:

• ChatHistory.store_chat_history_to_file() writes a file that is not UTF-8: a history containing héllo — 世界 lands on disk as h\xa8\xa6llo \xa1\xaa \xca\xc0\xbd\xe7 — GBK bytes for the same text. That file cannot be read on Linux/macOS or inside a container, where open() defaults to UTF-8: UnicodeDecodeError: 'utf-8' codec can't decode byte 0xa8, and any character outside the ANSI code page raises UnicodeEncodeError when writing.
• ChatHistory.load_chat_history_from_file() is the matching half, so the pair only round-trips on the machine that wrote it.
• OnnxGenAICompletionBase._load_genai_settings reads genai_config.json from a model directory without an encoding — those configs come from the model package and are UTF-8.

Change

-        with open(file_path, "w") as local_file:
+        with open(file_path, "w", encoding="utf-8") as local_file:

-        with open(file_path) as file:
+        with open(file_path, encoding="utf-8") as file:

-            with open(json_gen_ai_config) as file:
+            with open(json_gen_ai_config, encoding="utf-8") as file:

Three lines, no behaviour change on POSIX (UTF-8 is already the default there).

Verification

Windows 11 / CPython 3.13, semantic_kernel installed editable from python/:

# before: file is not UTF-8
utf-8 decodable: False -> 'utf-8' codec can't decode byte 0xa8 in position 218

# after: file is UTF-8 and the content is byte-identical to the input
utf8 decode ok: True
text matches original: True
round trip matches: True

python/tests/unit/contents/test_chat_history.py: 76 passed before and after (the suite does not pin an encoding, so it does not notice the difference — the fix is about interoperability, not about a failing test).

Scope

Only python/semantic_kernel/. A scan of the tree finds 32 more text-mode opens without an encoding, but they are in tests, samples and scripts; I kept this PR to the three in the shipped package so it stays reviewable in one pass. Happy to send the test/sample sweep separately if you want it.

[Copilot]

🟡 Changes recommended

Existing tests would still pass if explicit UTF-8 encodings were removed from ONNX initialization and chat-history persistence, so targeted regression coverage is needed.

2 open issues

• Test explicit UTF-8 encoding for ONNX config reads · New
• Test UTF-8 persistence and decoding of non-ASCII chat history · New

What changed in this PR

This PR makes file encoding explicit in the Python package, improving chat-history portability and ensuring ONNX model configuration files are read as UTF-8.

Changes:

• Use UTF-8 when writing and reading chat-history files.
• Use UTF-8 when reading ONNX genai_config.json files.

File	Description
python/​semantic_kernel/​contents/​chat_history.py	Specifies UTF-8 for history file reads and writes.
python/​semantic_kernel/​connectors/​ai/​onnx/​services/​onnx_gen_ai_completion_base.py	Specifies UTF-8 for model configuration reads.

_{Review effort: Balanced}

[Hyacinth-of-Security (lxcxjxhx)]

Added in 3fb8d30: test_onnx_config_is_read_as_utf8 in python/tests/unit/connectors/ai/onnx/services/test_onnx_chat_completion.py.

It filters open.call_args_list down to the genai_config.json call and asserts call.kwargs.get('encoding') == 'utf-8' on every one of them, so removing the explicit encoding fails the assertion rather than passing silently.

[Hyacinth-of-Security (lxcxjxhx)]

Added in 3fb8d30: test_chat_history_file_round_trip_keeps_non_ascii in python/tests/unit/contents/test_chat_history.py.

It stores héllo — 世界 and then asserts on bytes: 'héllo — 世界'.encode() in file_path.read_bytes(). Reading through read_text() would round-trip under any single-byte codec and prove nothing; the byte comparison is what fails when the encoding='utf-8' on the write is removed on a host whose default is not UTF-8.

[Hyacinth-of-Security (lxcxjxhx)]

Both points were right, and the fix is in 3fb8d30.

Chat history (python/tests/unit/contents/test_chat_history.py):

• test_chat_history_file_io_requests_utf8_explicitly spies on builtins.open for the duration of a store/load pair and asserts every call was made with encoding="utf-8".
• test_chat_history_file_round_trip_keeps_non_ascii stores héllo — 世界, asserts the UTF-8 bytes are in the file, and loads it back.

The honest split between them: the round-trip test passes on a UTF-8-default host even without the source change — I checked by reverting the encodings — so it documents the user-visible outcome rather than pinning the fix. The spy test is the one that fails when the encoding= arguments are removed, which is exactly the gap you pointed at.

ONNX (python/tests/unit/connectors/ai/onnx/services/test_onnx_chat_completion.py): test_onnx_config_is_read_as_utf8 filters the recorded open calls down to genai_config.json and asserts encoding="utf-8".

Replayed the regression, with the source encodings removed:

chat -> 1 failed, 1 passed     (the spy test fails; the round trip still passes)
onnx -> 1 failed

With the encodings restored: python/tests/unit/contents/test_chat_history.py + python/tests/unit/connectors/ai/onnx → 113 passed (needs onnxruntime-genai, which I installed to run the ONNX unit tests rather than skip them).